[1/2] Trend report: GenAI market in 2025 The technological base of the Russian GenAI (AI column)
I read it the other day. RedMadRobot Report About the GenAI market in Russia 2025 year. The report itself came out weighty. n (50+ pages) And I decided to break it down into a couple of posts. It all starts with a beautiful picture with the ontology of the GenAI market, which is then decomposed into a linear list of topics like infrastructure and telecom, models? data and industry knowledge, professional tools, applications, services, and so on. I'm going to go along with that, but not just from this report.
Infrastructure under GenAI is limited - the latest GPUs (For example, NVIDIA H100) officially unavailable, their purchase through parallel imports is more expensive 30%. Companies are forced to use previous generations (A100, etc.)These are offered by local cloud providers. As a result, the main bet is on your computing resources: on-premise deployment dominates.
Large language models (LLM) Only Sber trains from scratch., the rest of the foreign financiers (predominantly Chinese models). Sber has GigaChat, Yandex has YandexGPT, and T-Bank has T-Pro. The general trend is to move from racing for model size to finding optimal performance. Instead of rampant parameter growth (which comes down to hardware limits) The emphasis shifts to “small LLM” – specialized smaller models, retrained for tasks. RAG is actively used, using vector databases for jasemantic search. By the way, the latest Sber model was released in the popular architecture Mixture-of-Experts. (MoE) Sparse “expert mixes” where multiple narrow models co-cover different domains. The MoE became popular after the DeepSeek model that came out earlier this year.
High-quality datasets – GenAI fuel – are also under restrictions. After many years of web-scraping, the available corpus of texts is almost exhausted: according to MTS AI, there is not enough human content for new breakthroughs. Many fresh texts on the Internet are themselves created by neural networks, and training on them can lead to quality degradation. Developers are looking for new approaches: they train models on video, audio, images, expand multimodal samples. Russian players form their own sets: for example, Sber and Yandex collected terabytes of Russian text from open sources for their LLM; universities and companies publish benchmarks and datasets. (For example, RuLM, Russian analogues of SuperGLUE, etc.). But access to a number of Western databases is limited, and the quality of Russian-language data often requires cleaning and markup.
An ecosystem of assistive technologies is emerging around LLM. Libraries like HuggingFace Transformers, Russian frameworks (DeepPavlov et al.)specialized vector databases for semantic search of knowledge. There are also aggregators of neural networks - directories of GenAI services for business. In the field of MLOps, solutions arise for configuring and monitoring large models. However, there are no uniform standards: companies use different tool stacks, making it difficult to integrate models into existing systems. Integrators come to the rescue: after the departure of major consultants (Accenture, PwC et al.) The role of advisers is taken by local players.
As a result, Russia’s GenAI infrastructure is developing autonomously, relying on local resources and brains. In continuation I will talk about specific examples of the use of technology and options for the further development of this industry.
#Engineering #AI #Software #Management #Economics